You are viewing a plain text version of this content. The canonical link for it is here.
Posted to yarn-issues@hadoop.apache.org by "Arun Suresh (JIRA)" <ji...@apache.org> on 2018/01/10 08:19:00 UTC
[jira] [Comment Edited] (YARN-7681) Scheduler should double-check
placement constraint before actual allocation is made
[ https://issues.apache.org/jira/browse/YARN-7681?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16319876#comment-16319876 ]
Arun Suresh edited comment on YARN-7681 at 1/10/18 8:18 AM:
------------------------------------------------------------
Apologize for the last minute comment :)
[~cheersyang], actually, was just wondering, instead of passing the allocationTags down to the ResouceCommitRequest, and performing the check in the FicaSchedulerApp, can we not do this check inside the CapacityScheduler#attemptAllocationOnNode method, before we do the tryCommit ? I somehow feel this should be something the Scheduler does, and should not concern the app implementation.
Although your patch should work functionally, I feel it is odd that we use the checkPending flag to decide if we should perform the {{PlacementConstraintsUtil.canSatisfyConstraints}} check.
was (Author: asuresh):
Apologize for the last minute comment :)
[~cheersyang], actually, was just wondering, instead of passing the allocationTags down to the ResouceCommitRequest, and performing the check in the FicaSchedulerApp, can we not do this check inside the CapacityScheduler#attemptAllocationOnNode method, before we do the tryCommit ? I somehow feel this should be something the Scheduler does, and should concern the app implementation.
Although your patch should work functionally, I feel it is odd that we use the checkPending flag to decide if we should perform the {{PlacementConstraintsUtil.canSatisfyConstraints}} check.
> Scheduler should double-check placement constraint before actual allocation is made
> -----------------------------------------------------------------------------------
>
> Key: YARN-7681
> URL: https://issues.apache.org/jira/browse/YARN-7681
> Project: Hadoop YARN
> Issue Type: Sub-task
> Components: RM, scheduler
> Reporter: Weiwei Yang
> Assignee: Weiwei Yang
> Attachments: YARN-7681-YARN-6592.001.patch, YARN-7681-YARN-6592.002.patch
>
>
> This JIRA is created based on the discussions under YARN-7612, see comments after [this comment|https://issues.apache.org/jira/browse/YARN-7612?focusedCommentId=16303051&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-16303051]. AllocationTagsManager maintains tags info that helps to make placement decisions at placement phase, however tags are changing along with container's lifecycle, so it is possible that the placement violates the constraints at the scheduling phase. Propose to add an extra check in the scheduler to make sure constraints are still satisfied during the actual allocation.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org